Research On The Recognition Of Chinese Predicate Head Words Based On Neural Network

The predicate head is the focus of the sentence.Through predicate head recognition,each part of the grammatical elements of the sentence can be analyzed,so as to construct an event knowledge map with the predicate head as the core,which is of great significance to the study of dynamic changes and tracking of events.Identifying predicate head needs to judge which is the center of the sentence,the traditional models are mainly used to identify the sequence of shallow Tagging each word of the sentence classification.Since the identification of the predicate head for a single sentence,the sentence so semantic information acquisition is very important.The main work of this paper is divided into the following two aspects.For the issue of access predicate head identifying semantic information in context,this paper presents a model based on the depth of learning Highway-Bi LSTM network.First,a multi-layer stacked Bi LSTM network is used to obtain abstract semantic dependent information with different granularities within each sentence.Secondly,this article introduces each layer in the Highway network connection model to enable the high-speed flow of semantic information between layers.There can be multiple verbs in a sentence,but there is only one predicate head,that is,the problem of the uniqueness of the predicate head.In response to this problem,the output path of the predicate head is planned through the constraint layer,so as to ensure that each output sentence contains only one predicate head.The experimental results show that the F1 value of this method reaches 80.42% on the Chinese predicate head data set,which effectively improves the experimental performance.With the increase of the length of the sentence,the semantic-dependent long distance become a major factor limiting the performance of the model.Aiming at the problem of obtaining the global semantic information of sentences,this paper proposes a deep learning model based on bounding box regression.This method first converts each sentence into an abstract representation with global semantic dependent features,which is called a feature map.Then,the text bounding box is generated from the feature map.The bounding box represents the possible abstract representations of the predicate head.During training,the use of multi-objective learning framework classification confidence and learning the bounding box position offset with respect to the real predicate words.This method combines the neural network algorithm and the border regression algorithm to make full use of the global semantic information of the sentence.The experimental results show that the method achieves good performance in the Chinese predicate head data set,and the F1 value reaches 80.78%.